DMSACNN: Deep Multiscale Attentional Convolutional Neural Network for EEG-Based Motor Decoding.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Accurate electroencephalogram (EEG) signal decoding is crucial for advancing brain-computer interfaces (BCI).
• Motor imagery and motor execution (MI/ME) tasks are key for BCI control but face challenges in decoding accuracy.
• Existing methods often struggle with limited temporal information utilization and suboptimal feature selection.

Purpose of the Study:

• To introduce DMSACNN, a deep multiscale attention convolutional neural network tailored for enhanced MI/ME-EEG decoding.
• To address the limitations of current methods in capturing temporal dynamics and selecting discriminative features.
• To improve the robustness and accuracy of BCI systems through advanced EEG signal processing.

Main Methods:

• DMSACNN employs a deep multiscale temporal feature extraction module to capture diverse temporal patterns.
• A spatial convolutional module is integrated to extract relevant spatial features from EEG data.
• A local and global feature fusion attention module combines information for highly discriminative spatiotemporal feature extraction.

Main Results:

• DMSACNN achieved high accuracies: 78.20% on BCI-IV-2a, 96.34% on High Gamma, and 70.90% on OpenBMI datasets.
• The model demonstrated superior performance compared to most existing state-of-the-art methods in hold-out analyses.
• These results validate the effectiveness of the proposed deep multiscale attention approach.

Conclusions:

• DMSACNN shows significant potential for robust and accurate BCI applications.
• The developed method offers a valuable solution for improving MI/ME-EEG decoding accuracy.
• This advancement can lead to more efficient and reliable BCI systems for various applications.